Darrick,
On 2019/12/24 13:28, Darrick J. Wong wrote:
>> [...]
>>>> +
>>>> +static int zonefs_get_zone_info(struct zonefs_zone_data *zd)
>>>> +{
>>>> + struct block_device *bdev = zd->sb->s_bdev;
>>>> + int ret;
>>>> +
>>>> + zd->zones = kvcalloc(blkdev_nr_zones(bdev->bd_disk),
>>>> + sizeof(struct blk_zone), GFP_KERNEL);
>>>
>>> Hmm, so one 64-byte blk_zone structure for each zone on the disk?
>>>
>>> I have a 14TB SMR disk with ~459,000x 32M zones on it. That's going to
>>> require a contiguous 30MB memory allocation to hold all the zone
>>> information. Even your 15T drive from the commit message will need a
>>> contiguous 3.8MB memory allocation for all the zone info.
>>>
>>> I wonder if each zone should really be allocated separately and then
>>> indexed with an xarray or something like that to reduce the chance of
>>> failure when memory is fragmented or tight.
>>>
>>> That could be subsequent work though, since in the meantime that just
>>> makes zonefs mounts more likely to run out of memory and fail. I
>>> suppose you don't hang on to the huge allocation for very long.
>>
>> No, this memory allocation is only for mount. It is dropped as soon as
>> all the zone file inodes are created. Furthermore, this allocation is a
>> kvalloc, not a kmalloc. So there is no memory continuity requirement.
>> This is only an array of structures and that is not used to do IOs for
>> the report zone itself.
>>
>> I debated trying to optimize (I mean reducing the mount temporary memory
>> use) by processing mount in small chunks of zones instead of all zones
>> in one go. I kept simple, but rather brutal, approach to keep the code
>> simple. This can be rewritten and optimized at any time if we see
>> problems appearing.
>
> <nod> vmalloc space is quite limited on 32-bit platforms, so that's the
> most likely place you'll get complaints.
Yes, agreed. But the main use case for host-managed zoned drives (HDDs
or SSDs) being enterprise servers, 32-bits arch are unlikely to be an
issue. So for now, if there is no strong opposition, I would like to
keep the initialization as it is and revisit later if problems are reported.
--
Damien Le Moal
Western Digital Research
